Improving UNIWARD distortion function via isotropic construction and hierarchical merging

Abstract

Distortion function is designed for evaluating the cost of modifications in adaptive steganography. UNIWARD is a successful and popular distortion scheme which achieves high performance both for spatial and JPEG images. In this paper, we analyze the UNIWARD scheme with some empirical rules of distortion function designation. Based on that we propose our scheme to improve UNIWARD distortion. In our scheme, we focus on the symmetric characteristic of UNIWARD, and suggest that not only use original wavelet filters but also their flippings to calculate sub-models of UNIWARD distortion to maintain its isotropic properties. Moreover, we design several schemes to merge sub-models, which could maintain its invariance regard to flipping or rotation and improve its security against steganalysis detection. Experimental results show our revised UNIWARD achieves better performance for spatial and JPEG image in comparison with original UNIWARD.